push new default settings

app.py

@@ -143,7 +143,7 @@ def init_models():
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     question_answerer = pipeline(
         "question-answering", model='sultan/BioM-ELECTRA-Large-SQuAD2-BioASQ8B',
-        device=device
+        device=device, handle_impossible_answer=True,
     )
     reranker = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2', device=device)
     # queryexp_tokenizer = AutoTokenizer.from_pretrained("doc2query/all-with_prefix-t5-base-v1")
@@ -213,16 +213,16 @@ st.markdown("""
 with st.expander("Settings (strictness, context limit, top hits)"):
     concat_passages = st.radio(
         "Concatenate passages as one long context?",
-        ('no', 'yes'))
+        ('yes', 'no'))
     support_all = st.radio(
         "Use abstracts and titles as a ranking signal (if the words are matched in the abstract then the document is more relevant)?",
-        ('yes', 'no'))
+        ('no', 'yes'))
     support_abstracts = st.radio(
         "Use abstracts as a source document?",
         ('yes', 'no', 'abstract only'))
     strict_lenient_mix = st.radio(
         "Type of strict+lenient combination: Fallback or Mix? If fallback, strict is run first then if the results are less than context_lim we also search lenient. Mix will search them both and let reranking sort em out",
-        ('fallback', 'mix'))
+        ('mix', 'fallback'))
     confidence_threshold = st.slider('Confidence threshold for answering questions? This number represents how confident the model should be in the answers it gives. The number is out of 100%', 0, 100, 1)
     use_reranking = st.radio(
         "Use Reranking? Reranking will rerank the top hits using semantic similarity of document and query.",
@@ -323,7 +323,7 @@ def run_query(query, progress_bar):
     progress_bar.progress(50)
     if concat_passages == 'yes':
         context = '\n---'.join(contexts)
-        model_results = qa_model(question=query, context=context, top_k=10)
+        model_results = qa_model(question=query, context=context, top_k=10, doc_stride=512 // 2, max_answer_len=128, max_seq_len=512)
     else:
         context = ['\n---\n'+ctx for ctx in contexts]
         model_results = qa_model(question=[query]*len(contexts), context=context)